SmolVLA — arrange chocolate (3 cam, 50 ep)
Fine-tuned from lerobot/smolvla_base on makermods/arrange_chocolate_20260827_145746.
Task: pick up 8 gold chocolate and place in gold tray.
Training
| Dataset | 50 episodes / 49,995 frames @ 30 fps, 3 cameras (front, top, wrist) |
| Robot | so_follower, 6-DoF |
| Steps | 25,000 (32.0 epochs), batch 64, seed 1000 |
| LR | 1e-4, 1,000 warmup, cosine decay to 2.5e-06 over 25,000 |
| Config | freeze_vision_encoder=true, train_expert_only=true, use_amp=false |
| Final | loss 0.028, grad-norm 0.469 |
| Hardware | RTX 4090, 20.1 GB, 1.41 s/step, ~9h50m |
Checkpoints
25 checkpoints at checkpoints/{step:06d}/, each with pretrained_model and
training_state (resumable). Tagged by step, e.g. revision="020000".
Loss converged around step 20,000: the 20K->25K block improved only 3.4%, and
checkpoints 21K-25K oscillate within noise. Suggested evaluation order:
025000 (fully annealed), 020000 (converged, fewer epochs), 016000 (overfitting control).
There is no validation split, so these losses measure fit to the training set only.
Camera keys
Uses real dataset keys (observation.images.front / top / wrist) with
rename_map: {} — loads directly in lerobot-eval, lerobot-rollout, and
policy_server.py with no checkpoint rewrite.